Speaker
Description
Tissue microstructure dictates organ function, therefore a rigorous characterization of tissue architecture is imperative for understanding healthy and pathological processes [1]. While Synchrotron X-ray Phase Contrast Imaging (X-PCI) offers high-resolution, high-contrast and non-destructive 3D imaging of cardiac microstructure [2], the massive data volumes and complexity of these acquisitions make manual clinical interpretation impractical. To address this, we develop an autonomous computational pipeline for the robust exploration and interpretation of 3D microstructural environments.
Our proposed approach introduces an unsupervised 3D patch-based machine learning pipeline designed to autonomously characterize unlabeled volumetric X-PCI data, as shown in Figure 1. To remove the effect of spatial positioning in our samples, we propose a rotationally invariant representation for each patch. This is achieved by extracting 2D slices across several orientations and features through an ImageNet-pretrained ResNet-50 backbone [3,4,5], which we then aggregate into a single global descriptor using order-invariant operations, specifically max and mean pooling. Subsequently, we apply UMAP [6] for dimensionality reduction, generating a unified latent manifold where the 3D patches naturally cluster according to their intrinsic microstructural similarity.
To illustrate this framework, Heart Transplant Rejection (HTR) is utilized as a clinical case study, representing a critical scenario where subtle microstructural deterioration precedes macroscopic organ failure. HTR diagnosis currently relies on the histological analysis of endomyocardial biopsies (EMBs) via 2D Light Microscopy, a process that inherently destroys native 3D tissue architecture and limits spatial context. Our results show the pipeline successfully distinguishes diverse cardiovascular tissue types, clustering together cardiomyocytes, endocardium, and fine vascular structures, based purely on morphological signatures. Furthermore, the model successfully isolated pathological regions, mapping HTR-associated lymphocytic infiltrations into distinct spatial clusters.
Figure 1 (in attachements): Schematic overview of the proposed computational pipeline, detailing the sequential stages from initial data preprocessing to microstructural feature extraction and latent space projection.
References
[1] M. Seabra, N. Skreb, H. Dejea et al., Actas del XLII CASEIB, 2024, ISBN: 978-84-09-67332-2
[2] I. Planinc, I. Ilic, H. Dejea et al., Transplant International 36:11046, 2023, doi:10.3389/ti.2023.11046
[3] K. He, X. Zhang, S. Ren et al., Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, doi:10.1109/CVPR.2016.90
[4] O. Russakovsky, J. Deng, H. Su et al., International Journal of Computer Vision 115:211-252, 2015, doi:10.1007/s11263-015-0816-y
[5] A. Paszke, S. Gross, F. Massa et al., Advances in Neural Information Processing Systems 32, 2019
[6] L. McInnes, J. Healy, J. Melville, arXiv preprint arXiv:1802.03426, 2018, doi:10.48550/arXiv.1802.03426